D3RM : A Discrete Denoising Diffusion Refinement Model for Piano Transcription

This is the source code of D3RM paper accepted in ICASSP 2025. Regarding the reproducement of the paper, please let me know your concerns and feel free to comment them in the Issues part.

Installation

git clone https://github.com/hanshounsu/d3rm.git
pip -r install requirements.txt

Current project is based on pytorch-lightning 2.5.0, and an older version of natten (natten==0.15.1). You need to manually download the previous version of natten package after installing requirements.txt. (check the requirements.txt file)

Model Download

• Pretrained NAR-HC baseline model [link]
• Pretrained D3RM model [link]

Place the pretrained D3RM model in ./checkpoints/pretrained/

Download MAESTRO

Download here [link]

Place the dataset folder inside ./data

Training the model

python3 main_cli.py fit -c ./configs/D3RM_cli.yaml

Inference

python3 main_cli.py test -c ./logs/{TARGET_EXPERIMENT_CONFIG_PATH} --ckpt_path {TARGET_EXPERIMENT_CHECKPOINT_PATH}

Test with checkpoint model

python3 main_cli.py test -c ./configs/D3RM_cli.yaml --ckpt_path ./checkpoints/pretrained/step=0170000-metric_note_with_offsets_f1=0.9464.ckpt

License

This project is licensed under The MIT License.

Citations

@misc{hskim2023d3rm,
      title={D3RM : A Discrete Denoising Diffusion Refinement Model for Piano Transcription},
      author={Hounsu Kim, Taegyun Kwon, Juhan Nam},
      year={2024},
      eprint={},
      archivePrefix={arXiv},
      primaryClass={cs.Sound}
}